🚀 go-pugleaf

[ This article mentions no chips, no brands, no benchmarks.  It is
  about fuzzy issues as viewed from a systems management perspective.
  It is also 190 lines...  ]


Like Barry says, hardware technology is NOT the issue at this stage.

The issues (that I can remember now :-) are:

Education.

	Vendors have a tendency to ``forget'' to tell customers about
	the other half of the cost of the machine they're selling: support.
	The way it should be is to buy a slave (aka sysprog) along with
	any machine you buy, but several things conspire against awareness
	of this:

	- Many machines are sold to people that run one application, which
	  could be the O.S. of the machine for all they care.  This means
	  salescritters talk to a lot of people who really just want a
	  turnkey black box.  This means they don't keep the support issue
	  in mind when speaking to customers in other kinds of ships, and
	  certainly never bring it up... "But Mr. Jones, you realize that
	  with 10000$ to spend you really should use a few thousand on
	  proper system support, but unfortunately we don't have anything
	  you could buy for less than 9000$.  For your own good, Goodbye".
	  No way.

	- PCs (more "turnkey" machines) have given people the impression that
	  they can run their own box and that keeping a system running is
	  something that can be relegated to (literally, at times) slave labour.

	- Infrastructure is invisible.  People have to be reminded that it
	  exists, that it is necessary, and to cooperate with it.  Think
	  about your relationship with your Government and its Tax authority,
	  then scale that down to your local shop.  Does the thought give
	  you a warm fuzzy feeling that your money is being well spent on
	  necessities?  Probably not.
	
Psychology.

	People don't like other people spending their money.  In particular
	when they feel they have no authority over the process, or if they,
	personally, don't get special attention for their needs when they
	want it.  It feels like paying something for nothing, if all you're
	paying for is infrastructure.  They start viewing the infrastructure
	provider as a "them" in an arms-length relationship and look for
	ways of improving the service they get.  This typically means
	channelling $$ from large infrastructure to local equipment/labour,
	which undermines the infrastructure support and worsens the situation
	for everyone dependent on it.  Viscious circle.

	Most of such money is of course spent on equipment.  This was the
	precendent of the vax-class-minis to sun-class-workstations movement
	of a few years ago; people were moving from 1mips shared computing
	to 1.5mips on their desk, preciously guarded (ITS MINE! ALL MINE!
	GET YOUR MITTS OFF MY CPU!!!).  The effect was to go from largish
	empires that wouldn't or couldn't move quickly (largely due to
	lack of mini-class products at the time, later also due to the size
	of capital investment required), to small fiefdoms.  But people
	thought they were happy, and it *is* their money after all.

	So, they bought workstations.  They didn't need workstations.
	(I'm making sweeping statements here to make a point, I'm quite aware
	there are exceptions).  And still, today, most people don't need a
	"high-performance micro" on their desk; what they need is to get a
	job done in a productive (computer) work environment.  Nowadays that
	reads "bitmapped screen, window system, good response time".  If your
	job requires high-bandwidth communication to where you physically
	sit to do your work, then maybe having your own CPU is a reasonable
	way of fulfilling that requirement, but for NO OTHER REASON.

	Nevertheless, workstations were bought in droves for all the wrong
	reasons: status symbol, use-it-or-lose-it funding, and "wow! a
	computer on my desk; neato, now I can ignore you!".

	The end result is a LOT of compute cycles being spent in idle loops.

Economy.

	What happens if you try to fulfill the basic requirements using the
	Display Station approach (the sole purpose of the computer on your
	desk is to manage the interface to you and compute-cheap tasks),
	instead of the Work Station approach (your desktop computer does
	everything)?

	The first given then is that you save a lot of money by providing
	fancy terminals instead of computers on people's desks.  To
	maintain the productivity requirement invariant, your pour this
	money into centralized resources: CPU, I/O, Printers, etc.

	Try a rough calculation with today's prices: a display station
	is about half the price of a workstation, say.  For a 5k$ workstation
	you get 2.5k$ to beef up the central facilities if you buy a terminal
	instead.  Multiply this by the number of workstations at a largish
	site.  Now add this amount onto the existing infrastructure funds,
	and as a result you get a significant increase in shared resources.

	The advantages include:

		- there are no private resources to waste.
		- each user has the potential to use 100% of the shared
		  resources when they need to.
	
	The disadvantage that is always brought up as a counter argument
	is ROBUSTNESS.  Well, surprise surprise, a distributed environment
	is just as fragile as a centralized one with the same functionality
	(that's the consensus around here after years of observation), but
	it is MUCH more complex.  Centralized environments can be made
	very robust if well thought out.
	
	In addition to the economy of resource scheduling, there are the
	usual economies of scale inherent in an un-fragmented community.
	One might actually be able to afford decent I/O subsystems which
	cost order-of-magnitude of the main CPU.  These are all the more
	important in the context of this kind of timesharing environment,
	because of the timesharing load AND because there will always
	be people that would be I/O bound on the workstation they could
	afford for their desk (e.g. when it starts thrashing due to a large,
	theoretically compute-bound, job).
	
	As long as the shared CPU isn't idle waiting for I/O, it is effective
	and efficient computing as far as system management is concerned.
	
	If people have special needs, they buy the capability (a disk,
	another cpu, whatever), and hand it to the shared facility to
	make it available to themselves.  "If you want to use 600MB online,
	then either pay for file storage or give us a disk and you'll get
	that space".  The important part is that the disk sits with all
	the other disks, and therefore receives better attention than if
	it was on your desk -- A/C machine room, ease of service, bits are
	bits so other file space can be found if your disk died and it is
	important enough, competent management and monitoring, etc.

	If I were in that kind of environment (and I am :-), I wouldn't
	care about which resources were being used as long as they are
	1) paid for, and 2) easy to manage.  Often 2) translates into a
	variation of "how close is the problem to where I am now?".
	System staff hates running around all over the place, both physically
	and online.

Quality of Service.

	Done right, central facilities can give users a nice warm fuzzy good
	feeling.  I think the trick is to give them on-site personal attention,
	which is where centralization usually falls short.  There are solutions
	to that; the one we use here is to tell people to hire their own
	support people (aka slave labour in the University parlance).  They
	then act as liaisons between the shared facility and the specific
	users they interact with and perhaps as application specialists for the
	entire facility.  They do no infrastructure-type support.  The result
	is no duplication of work, and good service.

	But, as Barry says, "who cares" [which scheme people choose].
	Certainly I wouldn't, except for collateral effects on the shared
	facility and hence the other users.  In this situation, people really
	do get what they deserve.

	Note that none of my comments said anything about central authority,
	only central resources and resource-management.  There is a large
	psychological difference between the two concepts.

Immature Software.

	To come full circle: this whole problem is caused by the nature
	of today's computers, their software to be precise.  Each box is
	self-contained and thinks it owns the world.  This is the wrong
	premise and I think we'll need a second UNIX-like (r)evolution to
	make people realize this.

	I'd like to see a situation where people buy

		1) the interface hardware they need (capital cost)
		2) a chunk of computing resources (capital cost)
		3) system support (software/environment maintenance)

	The hardware is tangible; it would sit on their desk and connect to
	the communications infrastructure.  Most people need a fancy terminal,
	and should not have to change their equipment unless the interface
	technology changes and they NEED the new capabilities (e.g. bw->colour,
	Audio I/O, 3D digitizer, Holographic displays).  Deprecated Sun3/50s
	will work fine if you just want a screen,mouse,keyboard combination.
	(Actually, the way OS's are going, running the interface is about all
	they can do anyway :-).

	The chunk of resources is a capital cost contribution to a large
	shared "system", to be allocated within that system according to
	global needs.  That system in turn is tiered appropriately, using
	older slower technology (which is pretty fast these days) to
	control non-compute resources.  These days the depreciation period
	for hardware is 3 years, which means there is a lot of cheap capable
	computers out there, or there will be 3 years from now ;-).
	In a good system, they would be a reusable resource.


In large environments, the workstation concept is a fad.


					     rayan

		Artificial Intelligence/Numerical Analysis/Computational Theory
				    Shared Computing Facility
				      University of Toronto